llvm/test/CodeGen/AMDGPU/fptrunc.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
   2 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefixes=SI %s
   3 ; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-SAFE-SDAG %s
   4 ; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL,VI-SAFE-GISEL %s
   5 ; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-UNSAFE-SDAG %s
   6 ; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL,VI-UNSAFE-GISEL %s
   7 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-SAFE-SDAG %s
   8 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-SAFE-GISEL %s
   9 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-UNSAFE-SDAG %s
  10 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-UNSAFE-GISEL %s
  11 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-SAFE-SDAG %s
  12 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-SAFE-GISEL %s
  13 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-UNSAFE-SDAG %s
  14 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL %s
  15
  16 define amdgpu_kernel void @fptrunc_f64_to_f32(ptr addrspace(1) %out, double %in) {
  17 ; SI-LABEL: fptrunc_f64_to_f32:
  18 ; SI:       ; %bb.0:
  19 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
  20 ; SI-NEXT:    s_mov_b32 s7, 0xf000
  21 ; SI-NEXT:    s_mov_b32 s6, -1
  22 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
  23 ; SI-NEXT:    s_mov_b32 s4, s0
  24 ; SI-NEXT:    s_mov_b32 s5, s1
  25 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  26 ; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
  27 ; SI-NEXT:    s_endpgm
  28 ;
  29 ; VI-SDAG-LABEL: fptrunc_f64_to_f32:
  30 ; VI-SDAG:       ; %bb.0:
  31 ; VI-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
  32 ; VI-SDAG-NEXT:    s_mov_b32 s7, 0xf000
  33 ; VI-SDAG-NEXT:    s_mov_b32 s6, -1
  34 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  35 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  36 ; VI-SDAG-NEXT:    s_mov_b32 s4, s0
  37 ; VI-SDAG-NEXT:    s_mov_b32 s5, s1
  38 ; VI-SDAG-NEXT:    buffer_store_dword v0, off, s[4:7], 0
  39 ; VI-SDAG-NEXT:    s_endpgm
  40 ;
  41 ; VI-GISEL-LABEL: fptrunc_f64_to_f32:
  42 ; VI-GISEL:       ; %bb.0:
  43 ; VI-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
  44 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  45 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  46 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
  47 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
  48 ; VI-GISEL-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  49 ; VI-GISEL-NEXT:    s_endpgm
  50 ;
  51 ; GFX10-SDAG-LABEL: fptrunc_f64_to_f32:
  52 ; GFX10-SDAG:       ; %bb.0:
  53 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
  54 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  55 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  56 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
  57 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
  58 ; GFX10-SDAG-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  59 ; GFX10-SDAG-NEXT:    s_endpgm
  60 ;
  61 ; GFX10-GISEL-LABEL: fptrunc_f64_to_f32:
  62 ; GFX10-GISEL:       ; %bb.0:
  63 ; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
  64 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  65 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  66 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
  67 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
  68 ; GFX10-GISEL-NEXT:    buffer_store_dword v0, off, s[0:3], 0
  69 ; GFX10-GISEL-NEXT:    s_endpgm
  70 ;
  71 ; GFX11-SDAG-LABEL: fptrunc_f64_to_f32:
  72 ; GFX11-SDAG:       ; %bb.0:
  73 ; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
  74 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  75 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  76 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
  77 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
  78 ; GFX11-SDAG-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
  79 ; GFX11-SDAG-NEXT:    s_nop 0
  80 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  81 ; GFX11-SDAG-NEXT:    s_endpgm
  82 ;
  83 ; GFX11-GISEL-LABEL: fptrunc_f64_to_f32:
  84 ; GFX11-GISEL:       ; %bb.0:
  85 ; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
  86 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  87 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
  88 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
  89 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
  90 ; GFX11-GISEL-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
  91 ; GFX11-GISEL-NEXT:    s_nop 0
  92 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  93 ; GFX11-GISEL-NEXT:    s_endpgm
  94   %result = fptrunc double %in to float
  95   store float %result, ptr addrspace(1) %out
  96   ret void
  97 }
  98
  99 define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) {
 100 ; SI-LABEL: fptrunc_f64_to_f16:
 101 ; SI:       ; %bb.0:
 102 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 103 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 104 ; SI-NEXT:    s_mov_b32 s2, -1
 105 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 106 ; SI-NEXT:    s_mov_b32 s0, s4
 107 ; SI-NEXT:    s_mov_b32 s1, s5
 108 ; SI-NEXT:    s_lshr_b32 s4, s7, 8
 109 ; SI-NEXT:    s_and_b32 s5, s7, 0x1ff
 110 ; SI-NEXT:    s_and_b32 s8, s4, 0xffe
 111 ; SI-NEXT:    s_or_b32 s4, s5, s6
 112 ; SI-NEXT:    s_cmp_lg_u32 s4, 0
 113 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 114 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 115 ; SI-NEXT:    s_bfe_u32 s4, s7, 0xb0014
 116 ; SI-NEXT:    v_readfirstlane_b32 s5, v0
 117 ; SI-NEXT:    s_sub_i32 s6, 0x3f1, s4
 118 ; SI-NEXT:    s_add_i32 s10, s4, 0xfffffc10
 119 ; SI-NEXT:    s_or_b32 s11, s8, s5
 120 ; SI-NEXT:    v_med3_i32 v0, s6, 0, 13
 121 ; SI-NEXT:    s_lshl_b32 s4, s10, 12
 122 ; SI-NEXT:    s_or_b32 s5, s11, 0x1000
 123 ; SI-NEXT:    v_readfirstlane_b32 s6, v0
 124 ; SI-NEXT:    s_or_b32 s4, s11, s4
 125 ; SI-NEXT:    s_lshr_b32 s6, s5, s6
 126 ; SI-NEXT:    v_lshl_b32_e32 v0, s6, v0
 127 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, s5, v0
 128 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
 129 ; SI-NEXT:    v_readfirstlane_b32 s5, v0
 130 ; SI-NEXT:    s_or_b32 s5, s6, s5
 131 ; SI-NEXT:    s_cmp_lt_i32 s10, 1
 132 ; SI-NEXT:    s_cselect_b32 s6, s5, s4
 133 ; SI-NEXT:    s_and_b32 s8, s6, 7
 134 ; SI-NEXT:    s_cmp_gt_i32 s8, 5
 135 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 136 ; SI-NEXT:    s_cmp_eq_u32 s8, 3
 137 ; SI-NEXT:    s_cselect_b64 s[8:9], -1, 0
 138 ; SI-NEXT:    s_lshr_b32 s6, s6, 2
 139 ; SI-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
 140 ; SI-NEXT:    s_or_b32 s4, s4, s5
 141 ; SI-NEXT:    s_cmp_lg_u32 s4, 0
 142 ; SI-NEXT:    s_addc_u32 s4, s6, 0
 143 ; SI-NEXT:    s_cmp_lt_i32 s10, 31
 144 ; SI-NEXT:    s_cselect_b32 s6, s4, 0x7c00
 145 ; SI-NEXT:    s_cmp_lg_u32 s11, 0
 146 ; SI-NEXT:    s_cselect_b64 s[4:5], -1, 0
 147 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 148 ; SI-NEXT:    s_cmpk_eq_i32 s10, 0x40f
 149 ; SI-NEXT:    v_mov_b32_e32 v1, s6
 150 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 151 ; SI-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 152 ; SI-NEXT:    s_cselect_b64 vcc, -1, 0
 153 ; SI-NEXT:    s_lshr_b32 s4, s7, 16
 154 ; SI-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 155 ; SI-NEXT:    s_and_b32 s4, s4, 0x8000
 156 ; SI-NEXT:    v_or_b32_e32 v0, s4, v0
 157 ; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
 158 ; SI-NEXT:    s_endpgm
 159 ;
 160 ; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 161 ; VI-SAFE-SDAG:       ; %bb.0:
 162 ; VI-SAFE-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 163 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 164 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 165 ; VI-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 166 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s0, s4
 167 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s4, s7, 8
 168 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s8, s4, 0xffe
 169 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s4, s7, 0x1ff
 170 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s6
 171 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 172 ; VI-SAFE-SDAG-NEXT:    s_mov_b32 s1, s5
 173 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 174 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 175 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
 176 ; VI-SAFE-SDAG-NEXT:    s_bfe_u32 s5, s7, 0xb0014
 177 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s6, s8, s4
 178 ; VI-SAFE-SDAG-NEXT:    s_sub_i32 s8, 0x3f1, s5
 179 ; VI-SAFE-SDAG-NEXT:    v_med3_i32 v0, s8, 0, 13
 180 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s6, 0x1000
 181 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s8, v0
 182 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s8, s4, s8
 183 ; VI-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v0, s8
 184 ; VI-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc, s4, v0
 185 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
 186 ; VI-SAFE-SDAG-NEXT:    s_add_i32 s10, s5, 0xfffffc10
 187 ; VI-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
 188 ; VI-SAFE-SDAG-NEXT:    s_lshl_b32 s5, s10, 12
 189 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s4, s8, s4
 190 ; VI-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 191 ; VI-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s10, 1
 192 ; VI-SAFE-SDAG-NEXT:    s_cselect_b32 s11, s4, s5
 193 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s8, s11, 7
 194 ; VI-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s8, 5
 195 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 196 ; VI-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s8, 3
 197 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[8:9], -1, 0
 198 ; VI-SAFE-SDAG-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
 199 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s8, s11, 2
 200 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u64 s[4:5], 0
 201 ; VI-SAFE-SDAG-NEXT:    s_addc_u32 s4, s8, 0
 202 ; VI-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s10, 31
 203 ; VI-SAFE-SDAG-NEXT:    s_cselect_b32 s8, s4, 0x7c00
 204 ; VI-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 205 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 s[4:5], -1, 0
 206 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 207 ; VI-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 208 ; VI-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s10, 0x40f
 209 ; VI-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 210 ; VI-SAFE-SDAG-NEXT:    v_mov_b32_e32 v1, s8
 211 ; VI-SAFE-SDAG-NEXT:    s_cselect_b64 vcc, -1, 0
 212 ; VI-SAFE-SDAG-NEXT:    s_lshr_b32 s4, s7, 16
 213 ; VI-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 214 ; VI-SAFE-SDAG-NEXT:    s_and_b32 s4, s4, 0x8000
 215 ; VI-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s4, v0
 216 ; VI-SAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 217 ; VI-SAFE-SDAG-NEXT:    s_endpgm
 218 ;
 219 ; VI-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 220 ; VI-SAFE-GISEL:       ; %bb.0:
 221 ; VI-SAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 222 ; VI-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 223 ; VI-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 224 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 225 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 226 ; VI-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 227 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 228 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 229 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 230 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 231 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 232 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 233 ; VI-SAFE-GISEL-NEXT:    s_movk_i32 s5, 0x7e00
 234 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 235 ; VI-SAFE-GISEL-NEXT:    s_sub_i32 s7, 1, s4
 236 ; VI-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s4, 12
 237 ; VI-SAFE-GISEL-NEXT:    s_max_i32 s7, s7, 0
 238 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s6, s2, s6
 239 ; VI-SAFE-GISEL-NEXT:    s_min_i32 s7, s7, 13
 240 ; VI-SAFE-GISEL-NEXT:    s_bitset1_b32 s2, 12
 241 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s8, s2, s7
 242 ; VI-SAFE-GISEL-NEXT:    s_lshl_b32 s7, s8, s7
 243 ; VI-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s7, s2
 244 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 245 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s8, s2
 246 ; VI-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 247 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s2, s6
 248 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 249 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 250 ; VI-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 251 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 252 ; VI-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 253 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 254 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 255 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 256 ; VI-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 257 ; VI-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 258 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 259 ; VI-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 260 ; VI-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 261 ; VI-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 262 ; VI-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 263 ; VI-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 264 ; VI-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 265 ; VI-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 266 ; VI-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 267 ; VI-SAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 268 ; VI-SAFE-GISEL-NEXT:    s_endpgm
 269 ;
 270 ; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 271 ; VI-UNSAFE-SDAG:       ; %bb.0:
 272 ; VI-UNSAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 273 ; VI-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 274 ; VI-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 275 ; VI-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 276 ; VI-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 277 ; VI-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 278 ; VI-UNSAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 279 ; VI-UNSAFE-SDAG-NEXT:    s_endpgm
 280 ;
 281 ; VI-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 282 ; VI-UNSAFE-GISEL:       ; %bb.0:
 283 ; VI-UNSAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 284 ; VI-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 285 ; VI-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 286 ; VI-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 287 ; VI-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 288 ; VI-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 289 ; VI-UNSAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 290 ; VI-UNSAFE-GISEL-NEXT:    s_endpgm
 291 ;
 292 ; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 293 ; GFX10-SAFE-SDAG:       ; %bb.0:
 294 ; GFX10-SAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 295 ; GFX10-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 296 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s4, s3, 0x1ff
 297 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s3, 8
 298 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s2, s4, s2
 299 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s4, s5, 0xffe
 300 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s2, 0
 301 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s2, -1, 0
 302 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s2
 303 ; GFX10-SAFE-SDAG-NEXT:    s_bfe_u32 s2, s3, 0xb0014
 304 ; GFX10-SAFE-SDAG-NEXT:    s_sub_i32 s5, 0x3f1, s2
 305 ; GFX10-SAFE-SDAG-NEXT:    s_addk_i32 s2, 0xfc10
 306 ; GFX10-SAFE-SDAG-NEXT:    v_med3_i32 v1, s5, 0, 13
 307 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 308 ; GFX10-SAFE-SDAG-NEXT:    s_lshl_b32 s7, s2, 12
 309 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s6, v1
 310 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s5
 311 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s5, s4, 0x1000
 312 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s7, s4, s7
 313 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s6, s5, s6
 314 ; GFX10-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v1, s6
 315 ; GFX10-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc_lo, s5, v0
 316 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
 317 ; GFX10-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 318 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 319 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 1
 320 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, s7
 321 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s6, s5, 7
 322 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s6, 5
 323 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s7, -1, 0
 324 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s6, 3
 325 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s6, -1, 0
 326 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s5, 2
 327 ; GFX10-SAFE-SDAG-NEXT:    s_or_b32 s6, s6, s7
 328 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 329 ; GFX10-SAFE-SDAG-NEXT:    s_addc_u32 s5, s5, 0
 330 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 31
 331 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 332 ; GFX10-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 333 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 s4, -1, 0
 334 ; GFX10-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s2, 0x40f
 335 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s4
 336 ; GFX10-SAFE-SDAG-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 337 ; GFX10-SAFE-SDAG-NEXT:    s_lshr_b32 s2, s3, 16
 338 ; GFX10-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 339 ; GFX10-SAFE-SDAG-NEXT:    s_and_b32 s2, s2, 0x8000
 340 ; GFX10-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 341 ; GFX10-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 342 ; GFX10-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, s5, v0, vcc_lo
 343 ; GFX10-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s2, v0
 344 ; GFX10-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 345 ; GFX10-SAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 346 ; GFX10-SAFE-SDAG-NEXT:    s_endpgm
 347 ;
 348 ; GFX10-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 349 ; GFX10-SAFE-GISEL:       ; %bb.0:
 350 ; GFX10-SAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 351 ; GFX10-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 352 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 353 ; GFX10-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 354 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 355 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 356 ; GFX10-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 357 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 358 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 359 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 360 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 361 ; GFX10-SAFE-GISEL-NEXT:    s_movk_i32 s5, 0x7e00
 362 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 363 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 364 ; GFX10-SAFE-GISEL-NEXT:    s_sub_i32 s6, 1, s4
 365 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s7, s2, 0x1000
 366 ; GFX10-SAFE-GISEL-NEXT:    s_max_i32 s6, s6, 0
 367 ; GFX10-SAFE-GISEL-NEXT:    s_lshl_b32 s9, s4, 12
 368 ; GFX10-SAFE-GISEL-NEXT:    s_min_i32 s6, s6, 13
 369 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s2, s9
 370 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s8, s7, s6
 371 ; GFX10-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s8, s6
 372 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s6, s7
 373 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 374 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s6, s8, s6
 375 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 376 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s6, s2
 377 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 378 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 379 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 380 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 381 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 382 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 383 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 384 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 385 ; GFX10-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 386 ; GFX10-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 387 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 388 ; GFX10-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 389 ; GFX10-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 390 ; GFX10-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 391 ; GFX10-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 392 ; GFX10-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 393 ; GFX10-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 394 ; GFX10-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 395 ; GFX10-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 396 ; GFX10-SAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 397 ; GFX10-SAFE-GISEL-NEXT:    s_endpgm
 398 ;
 399 ; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 400 ; GFX10-UNSAFE-SDAG:       ; %bb.0:
 401 ; GFX10-UNSAFE-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 402 ; GFX10-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 403 ; GFX10-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 404 ; GFX10-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 405 ; GFX10-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 406 ; GFX10-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 407 ; GFX10-UNSAFE-SDAG-NEXT:    buffer_store_short v0, off, s[0:3], 0
 408 ; GFX10-UNSAFE-SDAG-NEXT:    s_endpgm
 409 ;
 410 ; GFX10-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 411 ; GFX10-UNSAFE-GISEL:       ; %bb.0:
 412 ; GFX10-UNSAFE-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 413 ; GFX10-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 414 ; GFX10-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 415 ; GFX10-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 416 ; GFX10-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 417 ; GFX10-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 418 ; GFX10-UNSAFE-GISEL-NEXT:    buffer_store_short v0, off, s[0:3], 0
 419 ; GFX10-UNSAFE-GISEL-NEXT:    s_endpgm
 420 ;
 421 ; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 422 ; GFX11-SAFE-SDAG:       ; %bb.0:
 423 ; GFX11-SAFE-SDAG-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
 424 ; GFX11-SAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 425 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s4, s3, 0x1ff
 426 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s3, 8
 427 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s2, s4, s2
 428 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s4, s5, 0xffe
 429 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s2, 0
 430 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s2, -1, 0
 431 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
 432 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s2
 433 ; GFX11-SAFE-SDAG-NEXT:    s_bfe_u32 s2, s3, 0xb0014
 434 ; GFX11-SAFE-SDAG-NEXT:    s_sub_i32 s5, 0x3f1, s2
 435 ; GFX11-SAFE-SDAG-NEXT:    s_addk_i32 s2, 0xfc10
 436 ; GFX11-SAFE-SDAG-NEXT:    v_med3_i32 v1, s5, 0, 13
 437 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 438 ; GFX11-SAFE-SDAG-NEXT:    s_lshl_b32 s7, s2, 12
 439 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 440 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s6, v1
 441 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s4, s4, s5
 442 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 443 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s5, s4, 0x1000
 444 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s7, s4, s7
 445 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s6, s5, s6
 446 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 447 ; GFX11-SAFE-SDAG-NEXT:    v_lshlrev_b32_e64 v0, v1, s6
 448 ; GFX11-SAFE-SDAG-NEXT:    v_cmp_ne_u32_e32 vcc_lo, s5, v0
 449 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
 450 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 451 ; GFX11-SAFE-SDAG-NEXT:    v_readfirstlane_b32 s5, v0
 452 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s5, s6, s5
 453 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 1
 454 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, s7
 455 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 456 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s6, s5, 7
 457 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_gt_i32 s6, 5
 458 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s7, -1, 0
 459 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_eq_u32 s6, 3
 460 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s6, -1, 0
 461 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s5, s5, 2
 462 ; GFX11-SAFE-SDAG-NEXT:    s_or_b32 s6, s6, s7
 463 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 464 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s6, 0
 465 ; GFX11-SAFE-SDAG-NEXT:    s_addc_u32 s5, s5, 0
 466 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lt_i32 s2, 31
 467 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 468 ; GFX11-SAFE-SDAG-NEXT:    s_cmp_lg_u32 s4, 0
 469 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 s4, -1, 0
 470 ; GFX11-SAFE-SDAG-NEXT:    s_cmpk_eq_i32 s2, 0x40f
 471 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s4
 472 ; GFX11-SAFE-SDAG-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 473 ; GFX11-SAFE-SDAG-NEXT:    s_lshr_b32 s2, s3, 16
 474 ; GFX11-SAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 475 ; GFX11-SAFE-SDAG-NEXT:    s_and_b32 s2, s2, 0x8000
 476 ; GFX11-SAFE-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 9, v0
 477 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 478 ; GFX11-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, 0x7c00, v0
 479 ; GFX11-SAFE-SDAG-NEXT:    v_cndmask_b32_e32 v0, s5, v0, vcc_lo
 480 ; GFX11-SAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 481 ; GFX11-SAFE-SDAG-NEXT:    v_or_b32_e32 v0, s2, v0
 482 ; GFX11-SAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 483 ; GFX11-SAFE-SDAG-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 484 ; GFX11-SAFE-SDAG-NEXT:    s_nop 0
 485 ; GFX11-SAFE-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 486 ; GFX11-SAFE-SDAG-NEXT:    s_endpgm
 487 ;
 488 ; GFX11-SAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 489 ; GFX11-SAFE-GISEL:       ; %bb.0:
 490 ; GFX11-SAFE-GISEL-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
 491 ; GFX11-SAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 492 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s3, 0x1ff
 493 ; GFX11-SAFE-GISEL-NEXT:    s_bfe_u32 s4, s3, 0xb0014
 494 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s5, s3, 8
 495 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s6, s2
 496 ; GFX11-SAFE-GISEL-NEXT:    s_addk_i32 s4, 0xfc10
 497 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s5, s5, 0xffe
 498 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 499 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 1, 0
 500 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 501 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s5, s2
 502 ; GFX11-SAFE-GISEL-NEXT:    s_movk_i32 s5, 0x7e00
 503 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s2, 0
 504 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s5, s5, 0x7c00
 505 ; GFX11-SAFE-GISEL-NEXT:    s_sub_i32 s6, 1, s4
 506 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s7, s2, 0x1000
 507 ; GFX11-SAFE-GISEL-NEXT:    s_max_i32 s6, s6, 0
 508 ; GFX11-SAFE-GISEL-NEXT:    s_lshl_b32 s9, s4, 12
 509 ; GFX11-SAFE-GISEL-NEXT:    s_min_i32 s6, s6, 13
 510 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s2, s9
 511 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s8, s7, s6
 512 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 513 ; GFX11-SAFE-GISEL-NEXT:    s_lshl_b32 s6, s8, s6
 514 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lg_u32 s6, s7
 515 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 516 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
 517 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s6, s8, s6
 518 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_lt_i32 s4, 1
 519 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s6, s2
 520 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s2, 7
 521 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s2, s2, 2
 522 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_eq_u32 s6, 3
 523 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s7, 1, 0
 524 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s6, 5
 525 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s6, 1, 0
 526 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 527 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s6, s7, s6
 528 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s6, s6, 1
 529 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 530 ; GFX11-SAFE-GISEL-NEXT:    s_add_i32 s2, s2, s6
 531 ; GFX11-SAFE-GISEL-NEXT:    s_cmp_gt_i32 s4, 30
 532 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, 0x7c00, s2
 533 ; GFX11-SAFE-GISEL-NEXT:    s_cmpk_eq_i32 s4, 0x40f
 534 ; GFX11-SAFE-GISEL-NEXT:    s_cselect_b32 s2, s5, s2
 535 ; GFX11-SAFE-GISEL-NEXT:    s_lshr_b32 s3, s3, 16
 536 ; GFX11-SAFE-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 537 ; GFX11-SAFE-GISEL-NEXT:    s_and_b32 s3, s3, 0x8000
 538 ; GFX11-SAFE-GISEL-NEXT:    s_or_b32 s2, s3, s2
 539 ; GFX11-SAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 540 ; GFX11-SAFE-GISEL-NEXT:    v_mov_b32_e32 v0, s2
 541 ; GFX11-SAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 542 ; GFX11-SAFE-GISEL-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 543 ; GFX11-SAFE-GISEL-NEXT:    s_nop 0
 544 ; GFX11-SAFE-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 545 ; GFX11-SAFE-GISEL-NEXT:    s_endpgm
 546 ;
 547 ; GFX11-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16:
 548 ; GFX11-UNSAFE-SDAG:       ; %bb.0:
 549 ; GFX11-UNSAFE-SDAG-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
 550 ; GFX11-UNSAFE-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 551 ; GFX11-UNSAFE-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 552 ; GFX11-UNSAFE-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 553 ; GFX11-UNSAFE-SDAG-NEXT:    s_mov_b32 s2, -1
 554 ; GFX11-UNSAFE-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 555 ; GFX11-UNSAFE-SDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 556 ; GFX11-UNSAFE-SDAG-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 557 ; GFX11-UNSAFE-SDAG-NEXT:    s_nop 0
 558 ; GFX11-UNSAFE-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 559 ; GFX11-UNSAFE-SDAG-NEXT:    s_endpgm
 560 ;
 561 ; GFX11-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16:
 562 ; GFX11-UNSAFE-GISEL:       ; %bb.0:
 563 ; GFX11-UNSAFE-GISEL-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
 564 ; GFX11-UNSAFE-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 565 ; GFX11-UNSAFE-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[2:3]
 566 ; GFX11-UNSAFE-GISEL-NEXT:    s_mov_b32 s2, -1
 567 ; GFX11-UNSAFE-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 568 ; GFX11-UNSAFE-GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 569 ; GFX11-UNSAFE-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
 570 ; GFX11-UNSAFE-GISEL-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 571 ; GFX11-UNSAFE-GISEL-NEXT:    s_nop 0
 572 ; GFX11-UNSAFE-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 573 ; GFX11-UNSAFE-GISEL-NEXT:    s_endpgm
 574   %result = fptrunc double %in to half
 575   %result_i16 = bitcast half %result to i16
 576   store i16 %result_i16, ptr addrspace(1) %out
 577   ret void
 578 }
 579
 580 define amdgpu_kernel void @fptrunc_v2f64_to_v2f32(ptr addrspace(1) %out, <2 x double> %in) {
 581 ; SI-LABEL: fptrunc_v2f64_to_v2f32:
 582 ; SI:       ; %bb.0:
 583 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0xd
 584 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 585 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 586 ; SI-NEXT:    s_mov_b32 s2, -1
 587 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 588 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 589 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 590 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 591 ; SI-NEXT:    s_endpgm
 592 ;
 593 ; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 594 ; VI-SDAG:       ; %bb.0:
 595 ; VI-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 596 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 597 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 598 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 599 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 600 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 601 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 602 ; VI-SDAG-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 603 ; VI-SDAG-NEXT:    s_endpgm
 604 ;
 605 ; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 606 ; VI-GISEL:       ; %bb.0:
 607 ; VI-GISEL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 608 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 609 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 610 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 611 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 612 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 613 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 614 ; VI-GISEL-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 615 ; VI-GISEL-NEXT:    s_endpgm
 616 ;
 617 ; GFX10-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 618 ; GFX10-SDAG:       ; %bb.0:
 619 ; GFX10-SDAG-NEXT:    s_clause 0x1
 620 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 621 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 622 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 623 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 624 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 625 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 626 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 627 ; GFX10-SDAG-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 628 ; GFX10-SDAG-NEXT:    s_endpgm
 629 ;
 630 ; GFX10-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 631 ; GFX10-GISEL:       ; %bb.0:
 632 ; GFX10-GISEL-NEXT:    s_clause 0x1
 633 ; GFX10-GISEL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 634 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 635 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 636 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 637 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 638 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 639 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 640 ; GFX10-GISEL-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 641 ; GFX10-GISEL-NEXT:    s_endpgm
 642 ;
 643 ; GFX11-SDAG-LABEL: fptrunc_v2f64_to_v2f32:
 644 ; GFX11-SDAG:       ; %bb.0:
 645 ; GFX11-SDAG-NEXT:    s_clause 0x1
 646 ; GFX11-SDAG-NEXT:    s_load_b128 s[4:7], s[0:1], 0x34
 647 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 648 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 649 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 650 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 651 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 652 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 653 ; GFX11-SDAG-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
 654 ; GFX11-SDAG-NEXT:    s_nop 0
 655 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 656 ; GFX11-SDAG-NEXT:    s_endpgm
 657 ;
 658 ; GFX11-GISEL-LABEL: fptrunc_v2f64_to_v2f32:
 659 ; GFX11-GISEL:       ; %bb.0:
 660 ; GFX11-GISEL-NEXT:    s_clause 0x1
 661 ; GFX11-GISEL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x34
 662 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 663 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 664 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 665 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 666 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 667 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 668 ; GFX11-GISEL-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
 669 ; GFX11-GISEL-NEXT:    s_nop 0
 670 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 671 ; GFX11-GISEL-NEXT:    s_endpgm
 672   %result = fptrunc <2 x double> %in to <2 x float>
 673   store <2 x float> %result, ptr addrspace(1) %out
 674   ret void
 675 }
 676
 677 define amdgpu_kernel void @fptrunc_v3f64_to_v3f32(ptr addrspace(1) %out, <3 x double> %in) {
 678 ; SI-LABEL: fptrunc_v3f64_to_v3f32:
 679 ; SI:       ; %bb.0:
 680 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 681 ; SI-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x11
 682 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x15
 683 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 684 ; SI-NEXT:    s_mov_b32 s6, -1
 685 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 686 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[10:11]
 687 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[8:9]
 688 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[0:1]
 689 ; SI-NEXT:    buffer_store_dword v2, off, s[4:7], 0 offset:8
 690 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 691 ; SI-NEXT:    s_endpgm
 692 ;
 693 ; VI-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 694 ; VI-SDAG:       ; %bb.0:
 695 ; VI-SDAG-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x54
 696 ; VI-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x44
 697 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 698 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 699 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[2:3]
 700 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 701 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 702 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 703 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 704 ; VI-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 705 ; VI-SDAG-NEXT:    s_endpgm
 706 ;
 707 ; VI-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 708 ; VI-GISEL:       ; %bb.0:
 709 ; VI-GISEL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 710 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 711 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 712 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 713 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 714 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 715 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 716 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 717 ; VI-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 718 ; VI-GISEL-NEXT:    s_endpgm
 719 ;
 720 ; GFX10-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 721 ; GFX10-SDAG:       ; %bb.0:
 722 ; GFX10-SDAG-NEXT:    s_clause 0x2
 723 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x54
 724 ; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x44
 725 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 726 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 727 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[2:3]
 728 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 729 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 730 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 731 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 732 ; GFX10-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 733 ; GFX10-SDAG-NEXT:    s_endpgm
 734 ;
 735 ; GFX10-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 736 ; GFX10-GISEL:       ; %bb.0:
 737 ; GFX10-GISEL-NEXT:    s_clause 0x1
 738 ; GFX10-GISEL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 739 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 740 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 741 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 742 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 743 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 744 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 745 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 746 ; GFX10-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
 747 ; GFX10-GISEL-NEXT:    s_endpgm
 748 ;
 749 ; GFX11-SDAG-LABEL: fptrunc_v3f64_to_v3f32:
 750 ; GFX11-SDAG:       ; %bb.0:
 751 ; GFX11-SDAG-NEXT:    s_clause 0x2
 752 ; GFX11-SDAG-NEXT:    s_load_b64 s[2:3], s[0:1], 0x54
 753 ; GFX11-SDAG-NEXT:    s_load_b128 s[4:7], s[0:1], 0x44
 754 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 755 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 756 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[2:3]
 757 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 758 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 759 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 760 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 761 ; GFX11-SDAG-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], 0
 762 ; GFX11-SDAG-NEXT:    s_nop 0
 763 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 764 ; GFX11-SDAG-NEXT:    s_endpgm
 765 ;
 766 ; GFX11-GISEL-LABEL: fptrunc_v3f64_to_v3f32:
 767 ; GFX11-GISEL:       ; %bb.0:
 768 ; GFX11-GISEL-NEXT:    s_clause 0x1
 769 ; GFX11-GISEL-NEXT:    s_load_b256 s[4:11], s[0:1], 0x44
 770 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 771 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 772 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 773 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 774 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 775 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 776 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 777 ; GFX11-GISEL-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], 0
 778 ; GFX11-GISEL-NEXT:    s_nop 0
 779 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 780 ; GFX11-GISEL-NEXT:    s_endpgm
 781   %result = fptrunc <3 x double> %in to <3 x float>
 782   store <3 x float> %result, ptr addrspace(1) %out
 783   ret void
 784 }
 785
 786 define amdgpu_kernel void @fptrunc_v4f64_to_v4f32(ptr addrspace(1) %out, <4 x double> %in) {
 787 ; SI-LABEL: fptrunc_v4f64_to_v4f32:
 788 ; SI:       ; %bb.0:
 789 ; SI-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x11
 790 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 791 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 792 ; SI-NEXT:    s_mov_b32 s2, -1
 793 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 794 ; SI-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 795 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 796 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 797 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 798 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 799 ; SI-NEXT:    s_endpgm
 800 ;
 801 ; VI-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 802 ; VI-SDAG:       ; %bb.0:
 803 ; VI-SDAG-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 804 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 805 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 806 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 807 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 808 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 809 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 810 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 811 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 812 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 813 ; VI-SDAG-NEXT:    s_endpgm
 814 ;
 815 ; VI-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 816 ; VI-GISEL:       ; %bb.0:
 817 ; VI-GISEL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 818 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 819 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 820 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 821 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 822 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 823 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 824 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 825 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 826 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 827 ; VI-GISEL-NEXT:    s_endpgm
 828 ;
 829 ; GFX10-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 830 ; GFX10-SDAG:       ; %bb.0:
 831 ; GFX10-SDAG-NEXT:    s_clause 0x1
 832 ; GFX10-SDAG-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 833 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 834 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 835 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 836 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 837 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 838 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 839 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 840 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 841 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 842 ; GFX10-SDAG-NEXT:    s_endpgm
 843 ;
 844 ; GFX10-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 845 ; GFX10-GISEL:       ; %bb.0:
 846 ; GFX10-GISEL-NEXT:    s_clause 0x1
 847 ; GFX10-GISEL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x44
 848 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 849 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 850 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 851 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 852 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 853 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 854 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 855 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 856 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 857 ; GFX10-GISEL-NEXT:    s_endpgm
 858 ;
 859 ; GFX11-SDAG-LABEL: fptrunc_v4f64_to_v4f32:
 860 ; GFX11-SDAG:       ; %bb.0:
 861 ; GFX11-SDAG-NEXT:    s_clause 0x1
 862 ; GFX11-SDAG-NEXT:    s_load_b256 s[4:11], s[0:1], 0x44
 863 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 864 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 865 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
 866 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 867 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 868 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 869 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 870 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 871 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
 872 ; GFX11-SDAG-NEXT:    s_nop 0
 873 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 874 ; GFX11-SDAG-NEXT:    s_endpgm
 875 ;
 876 ; GFX11-GISEL-LABEL: fptrunc_v4f64_to_v4f32:
 877 ; GFX11-GISEL:       ; %bb.0:
 878 ; GFX11-GISEL-NEXT:    s_clause 0x1
 879 ; GFX11-GISEL-NEXT:    s_load_b256 s[4:11], s[0:1], 0x44
 880 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 881 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
 882 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 883 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 884 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 885 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 886 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 887 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 888 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
 889 ; GFX11-GISEL-NEXT:    s_nop 0
 890 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 891 ; GFX11-GISEL-NEXT:    s_endpgm
 892   %result = fptrunc <4 x double> %in to <4 x float>
 893   store <4 x float> %result, ptr addrspace(1) %out
 894   ret void
 895 }
 896
 897 define amdgpu_kernel void @fptrunc_v8f64_to_v8f32(ptr addrspace(1) %out, <8 x double> %in) {
 898 ; SI-LABEL: fptrunc_v8f64_to_v8f32:
 899 ; SI:       ; %bb.0:
 900 ; SI-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x19
 901 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 902 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 903 ; SI-NEXT:    s_mov_b32 s2, -1
 904 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 905 ; SI-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 906 ; SI-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 907 ; SI-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 908 ; SI-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 909 ; SI-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
 910 ; SI-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
 911 ; SI-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
 912 ; SI-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
 913 ; SI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 914 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 915 ; SI-NEXT:    s_endpgm
 916 ;
 917 ; VI-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 918 ; VI-SDAG:       ; %bb.0:
 919 ; VI-SDAG-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x64
 920 ; VI-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 921 ; VI-SDAG-NEXT:    s_mov_b32 s3, 0xf000
 922 ; VI-SDAG-NEXT:    s_mov_b32 s2, -1
 923 ; VI-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 924 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
 925 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
 926 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
 927 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
 928 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 929 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 930 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 931 ; VI-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 932 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 933 ; VI-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 934 ; VI-SDAG-NEXT:    s_endpgm
 935 ;
 936 ; VI-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
 937 ; VI-GISEL:       ; %bb.0:
 938 ; VI-GISEL-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x64
 939 ; VI-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 940 ; VI-GISEL-NEXT:    s_mov_b32 s2, -1
 941 ; VI-GISEL-NEXT:    s_mov_b32 s3, 0xf000
 942 ; VI-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 943 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 944 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 945 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 946 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 947 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
 948 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
 949 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
 950 ; VI-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
 951 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 952 ; VI-GISEL-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 953 ; VI-GISEL-NEXT:    s_endpgm
 954 ;
 955 ; GFX10-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 956 ; GFX10-SDAG:       ; %bb.0:
 957 ; GFX10-SDAG-NEXT:    s_clause 0x1
 958 ; GFX10-SDAG-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x64
 959 ; GFX10-SDAG-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 960 ; GFX10-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
 961 ; GFX10-SDAG-NEXT:    s_mov_b32 s2, -1
 962 ; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 963 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
 964 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
 965 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
 966 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
 967 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 968 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 969 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 970 ; GFX10-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 971 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 972 ; GFX10-SDAG-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 973 ; GFX10-SDAG-NEXT:    s_endpgm
 974 ;
 975 ; GFX10-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
 976 ; GFX10-GISEL:       ; %bb.0:
 977 ; GFX10-GISEL-NEXT:    s_clause 0x1
 978 ; GFX10-GISEL-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x64
 979 ; GFX10-GISEL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 980 ; GFX10-GISEL-NEXT:    s_mov_b32 s2, -1
 981 ; GFX10-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
 982 ; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 983 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
 984 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
 985 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
 986 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
 987 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
 988 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
 989 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
 990 ; GFX10-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
 991 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 992 ; GFX10-GISEL-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16
 993 ; GFX10-GISEL-NEXT:    s_endpgm
 994 ;
 995 ; GFX11-SDAG-LABEL: fptrunc_v8f64_to_v8f32:
 996 ; GFX11-SDAG:       ; %bb.0:
 997 ; GFX11-SDAG-NEXT:    s_clause 0x1
 998 ; GFX11-SDAG-NEXT:    s_load_b512 s[4:19], s[0:1], 0x64
 999 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
1000 ; GFX11-SDAG-NEXT:    s_mov_b32 s3, 0x31016000
1001 ; GFX11-SDAG-NEXT:    s_mov_b32 s2, -1
1002 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
1003 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
1004 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
1005 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
1006 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
1007 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
1008 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
1009 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
1010 ; GFX11-SDAG-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
1011 ; GFX11-SDAG-NEXT:    s_clause 0x1
1012 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16
1013 ; GFX11-SDAG-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
1014 ; GFX11-SDAG-NEXT:    s_nop 0
1015 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
1016 ; GFX11-SDAG-NEXT:    s_endpgm
1017 ;
1018 ; GFX11-GISEL-LABEL: fptrunc_v8f64_to_v8f32:
1019 ; GFX11-GISEL:       ; %bb.0:
1020 ; GFX11-GISEL-NEXT:    s_clause 0x1
1021 ; GFX11-GISEL-NEXT:    s_load_b512 s[4:19], s[0:1], 0x64
1022 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
1023 ; GFX11-GISEL-NEXT:    s_mov_b32 s2, -1
1024 ; GFX11-GISEL-NEXT:    s_mov_b32 s3, 0x31016000
1025 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
1026 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
1027 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v1, s[6:7]
1028 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v2, s[8:9]
1029 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v3, s[10:11]
1030 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v4, s[12:13]
1031 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v5, s[14:15]
1032 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v6, s[16:17]
1033 ; GFX11-GISEL-NEXT:    v_cvt_f32_f64_e32 v7, s[18:19]
1034 ; GFX11-GISEL-NEXT:    s_clause 0x1
1035 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
1036 ; GFX11-GISEL-NEXT:    buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16
1037 ; GFX11-GISEL-NEXT:    s_nop 0
1038 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
1039 ; GFX11-GISEL-NEXT:    s_endpgm
1040   %result = fptrunc <8 x double> %in to <8 x float>
1041   store <8 x float> %result, ptr addrspace(1) %out
1042   ret void
1043 }